Unabridged Learning

	Rich Caruana	Joseph O'Sullivan

Lately we've been playing with a half-baked machine learning idea we
are calling "unabridged learning". Our goal is to make learning more
robust in the face of unexpected input data, by learning more complete
models from the available input features. In this talk, we will try to
define unabridged learning, present concrete examples of abridged vs
unabridged learning, and discuss methods for constructing unabridged
learners.  

Unabridged learning grew out of our research in inductive transfer.
We think it's an exciting new idea that looks promising, but we
haven't carried it very far yet.  At the seminar we'd like to incite
discussion as to whether this is an interesting approach, what
alternatives should be considered, and what learners already do some
form of unabridged learning.  AND, as an added bonus, we promise an
eat-n-park cookie to anyone who can come up with a better name than
"unabridged".